Professional Documents
Culture Documents
high performance
(and why they arent even better)
Josh Reiss
josh.reiss@elec.qmul.ac.uk
Centre for Digital Music
Queen Mary, University of London
UK
Approach
explain the basics of the signal-to-noise ratio (SNR),
theory and estimation in PCM and SDM converters
derive the formula for the SNR of an arbitrary SDM
Quantization Assumption
q=2V/(2b-1)
quantization error - difference
between input & output to quantiser
eq=Q(x)-x
rounding quantizer - assigns each
input to nearest quantization level
1.0
0.8
0.6
0.4
Q(x)
2/(2^3-1)=2/7
0.2
0.0
-0.2
-0.4
-0.6
-0.8
-1.0
-1.0
-0.8
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
0.8
1.0
q / 2 eq (n) q / 2
The allowed values in the output signal, after quantization, are called quantization
levels, whereas the distance between 2 successive levels, q, is called the
quantization step size. Consider a quantiser with b bits covering the range 2V, from
+V to -V. Then there are 2b quantization levels, and the width of each quantization
step is
q=2V/(2b-1)
This is depicted for a 3 bit quantiser and V=1.
The rounding quantizer assigns each input sample x(n) to the nearest quantization
level. The quantization error is simply the difference between the input and output to
the quantiser, eq=Q(x)-x. It can easily be seen that the quantization error eq(n) is
always bounded by +/- q/2
1/ q | eq | q / 2
p (eq ) =
0 | eq |> q / 2
p(eq)
1/q
not exact
q=2V/(2b-1)
-q/2
+q/2
eq
eq = E{eq } =
q/2
eq p (eq )deq =
= E{(eq eq ) } = E{eq } =
2
e
q / 2
Or
e2 =
eq
1
deq = 0
q
q/2
eq p (eq ) deq =
2
eq 2
q / 2
1
q2
deq =
q
12
12
12
3(2b -1) 2 3 22b
x2 = E{(ex ex ) 2 } = E{ex 2 } =
1
A2
2
(
cos(2
/
))
A
t
T
dt
=
2
T 0
If the sampling rate satisfies the sampling theorem, fs>2fB, quantization is the only
error in the A/D conversion process (jitter and other effects are not considered
here). Using the assumption of uniform distribution, we can find the average
quantization noise and the quantization noise power.
Quantisation noise power is found using textbook definitions
To find the SNR, we also need to estimate the signal power. Now assume we are
quantizing a sinusoidal signal of amplitude A, x(t)=Acos(2pt/T). The average power
of the signal is thus A2/2
From and , the signal-to-noise ratio may now be given by,
Thus the signal-to-noise ratio increases by approximately 6dB for every bit in the
quantiser. Using this formula, an audio signal encoded onto CD (a 16 bit format)
using PCM, has a maximum SNR of 98.08. Also note from , that the SNR is linearly
related to the signal strength in decibels.
Signal power
signal-to-noise ratio,
SNR(dB)=20log10
e2 =
V
V
3(2b -1) 2 3 2 2b
x2 = A2 / 2
x
A
20log10 + 6.02b + 1.76
e
V
Oversampling
Previously
highest
assumed
input frequency was fB, signal acquired at Nyquist rate, fs=2fB.
decreases
pe(f)
Nyquist rate PCM, f s/2=f B
4x Oversampled PCM, f s/2=4f B
8x Oversampled PCM, f s/2=8f B
fB
4fB
8fB
2f 1
q2
n2 = eq 2 p (eq ) deq = eq 2 B deq = r
= e2 / OSR
f
q
2
12
q / 2
Most noise power is now located outside of signal band
signal power occurs over signal band only
remains unchanged
The above discussion assumed that the highest possible frequency in the input
signal was some value fB and that the signal is acquired at the Nyquist rate, fs=2fB.
However, lets now assume that the signal is oversampled such that the rate is
fs=2r+1fB, that is, the oversampling ratio is OSR=2r=fs/2fB. Thus, the quantization
noise is spread over a larger frequency range yet we are only concerned with noise
below the Nyquist frequency.
The in-band quantization noise power is now given by the total quantisation noise
power divided by the OSR.
Most of the noise power is now located outside of the signal band. The signal power
occurs over the signal band only, so it remains unchanged.
The signal-to-noise ratio may now be given by, the previous SNR +10log10OSR
Thus for every doubling of the oversampling ratio, the SNR improves by 3dB. The
6dB improvement with each bit in the quantiser remains, so we can say that
doubling the oversampling ratio increases the effective number of bits by a bit.
n2 =
fB
Se ( f ) | NTF ( f ) |2 df =
fB
e2
fs
fB
| NTF ( f ) |
df
fB
=
2
n
e2
fs
fB
fB
1df =
2 e2 f B
= e2 / OSR
fs
Same as before.
Equivalent of 16 bit Nyquist rate PCM (98dB SNR)
8bit 2.64GHz
As a lead-in to the following sections which deal with sigma delta modulation,
consider that the (in-band) quantization noise power can also be given as a function
of the Noise Transfer Function and the power spectral density of the (unshaped)
quantization noise. Since there is no noise shaping, just oversampling, we have that
the noise transfer function is uniformly one over the range [-fB,fB]. So we get the
same as .before
Oversampling
Doubling the oversampling ratio results in
approximately a 3dB drop in the noise floor
10
X(n)+
U(n+1)
+
U(n)
Q
Delay
Q(n)
X(z) +
+
-
Y(z)
-1
z
1-z-1
given as
11
X(n)+
U(n+1)
+
U(n)
Q
Delay
X(z) +
Q(n)
Y ( z ) = X ( z ) z 1 + E ( z )(1 z 1 )
Y(z)
+
z-1
1-z-1
So,
NTF ( f ) = 1 e j 2 f / fs
| NTF ( f ) |2 = ... = 4sin 2 ( f / f s )
=
2
n
fB
fB
Se ( f ) | NTF ( f ) | df =
2
e2
fs
fB
4sin 2 ( f / f s )df
fB
Here, Se(f) is the power spectral density of the unshaped quantization noise
12
3 OSR
SNR(dB)
10log10
x2
A
20log10 + 6.02b + 9.03r 3.41
2
V
n
The effect of first order noise shaping is evident. We now get an improvement of
9dB for each doubling of the oversampling ratio, rather than the 3dB improvement
which occurs without noiseshaping.
13
+
+
Y(z)
z-1
+
-
z-1
Transfer function
n2 =
e2
fs
fB
[2sin( f / f )] df
4
fB
4
5 OSR
e2
14
4
5 OSR
e2
SNR (dB)
x2
A
20log10 + 6.02b + 15.05r 11.14
V
n2
15dB improvement in the SNR with each doubling of the
oversampling ratio
Equivalent of 16 bit Nyquist rate PCM (98dB SNR)
10log10
1bit 6.12MHz
Compared with the 1st order SDM, this provides more suppression of the
quantization noise over the low frequencies, and more amplification of the noise
outside the signal band.
Thus we see a large improvement in moving to a second order SDM. There is now
a 15dB improvement in the SNR with each doubling of the oversampling ratio.
15
Magnitude Spectrum
7
6
5
4
3
2
1
0
20
40
60
80
f (kHz)
100
120
140
160
16
n2 =
Giving
fs
fB
[2sin( f / f s )]N df e2
fB
SNR(dB)=10log10
20log10
e2
2N
(2 N + 1)OSR 2 N +1
x2
(2 N + 1)2(2 N +1) r
+ 10log10
2
e
2N
A
+ 6.02b + 1.76 + 10log10 (2 N + 1) 9.94N + 3.01(2 N + 1)r
V
17
this is an approximation
depends on
coefficients of the modulator
on approximations used in derivation
Assume high oversampling ratio
High number of bits in quantiser
other factors.
18
SNR vs Bits
120
Predicted from
20log10
A
+ 6.02b + 1.76
V
assumes
V2
V2
=
b
2
3(2 -1)
3 22b
2
e
110
x
e
100
Simulated SNR
Predicted SNR
90
SNR (dB)
SNR(dB)=20log10
80
70
60
50
40
30
20
10
2
10
12
# of Bits
14
16
18
20
19
250
240
230
SNR
220
e2 =
V
3(2b -1) 2
210
200
190
Input Amplitude
0.1
0.6
0.2
0.7
0.3
0.8
0.4
0.9
0.5
1.0
180
170
160
150
6
7
8
9
Number of Bits
10
11
12
13
14
15
16
20
200
SNR (dB)
220
180
160
140
3. uniform noise
distribution
constant error between
theory & simulation for PCM
120
100
5
6
log2(OSR)
10
21
Example
64 times oversampled 1-bit A/D converter, using 5th order SDM
SNR(dB) 20log10
A
+ 167.15
V
No 5th order, 1 bit SDM gives such high SNR for 64 times OSR
22
Shameless promotion
AES 31st International Conference
23